At the Convergence of Input and Process Models of Group Discussion: A Comparison of Participation Rates across Time, Persons, and Groups

Abstract

We investigate the stability and change of participation patterns in small groups by examining two longitudinal data sets at the individual and group levels of analysis. Rejecting the dichotomy between input and process models, we advance a view at the convergence of these two perspectives. We argue that stability in participation reflects input factors and that change emerges from process mechanisms. Study 1 analyzed discussion data from zero-history laboratory groups that worked on three similar tasks in succession, each with stable membership across the tasks. Results showed significant variation within participants and between groups, indicating that group members varied their participation as needed and that group-level factors influenced participation. Study 2 analyzed longitudinal data collected from the Australian Citizens' Parliament, where tasks and group membership varied over time. Study 2 replicated Study 1's findings, but analyses showed more complex patterns of both stability and change across groups and tasks. Taken together, results from the two studies support our position that both input and process mechanisms cause variation in participation. Our Conclusion examines how structural features and participation impact democratic group deliberation.

Keywords:

• Participation
• Deliberation
• Group Interaction
• Multilevel Modeling
• Decision making

Acknowledgments

The data in Study 2 could not have been collected and transcribed effectively without the diligent oversight of Ron Lubensky. Financial support for the research presented herein came from the Australian Research Council (a generous ARC-Linkage grant—No. LP0882714), as well as from the US National Science Foundation (NSF) Directorate for Social, Behavioral and Economic Sciences (Political Science and Decision, Risk and Management Sciences Programs, Grant No. 0908554). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of ARC or NSF. We hope that this paper honors the memory of Renee Meyers, who co-authored this paper before passing away on March 16, 2012.

Notes

[1] The scenarios did not provide a “0 in 10” option, as it would be irrational to advise the man to board the flight if pain were guaranteed to return.

[2] The “latent group model” on which we base this analysis is actually a three-level implementation, with the first level containing the so-called “switching regression” that identifies the task; the second level is the individual, and; the third is the group. The residual in this model is set to zero. Each level of analysis has six random parameters, three for the variances associated with each task and three covariances among the tasks. This covariance matrix is often referred to as “unstructured” because no constraints are placed on estimation. The “variance components” matrix, for example, estimates only the variances and not the covariances for the set of given random variables.

[3] In standard growth models, the intercept is the mean at T1, and the random variance component estimates the heterogeneity of the individual participation rates relative to the mean. This would be our test of H1 under normal circumstances. Here, because task order was randomized (and order not recorded), it is pointless to estimate both the intercept and associated random variance component. Moreover, under normal circumstances, we would test H2 by evaluating fixed-effects slopes and their random variance components, which provide evidence of covariation within and across individuals. Such a test is not possible here.

[4] Correlations are computed from the covariance matrix in the usual fashion by taking the ratio of the relevant covariance to the square root of the product of the relevant variances at the desired level of analysis. Significance levels are derived from the relevant covariance.

[5] The ICC is computed by dividing the group variance for a given task by the sum of the group and individual variances for that task. Significance values are derived from the relevant group-level variance.

[6] The chi-squares shown herein test the significance of overall model fit, based on the decrease in the model deviance score, with the degrees of freedom equal to the number of new parameters added to the model.

[7] Prior to attending the Canberra meetings, the vast majority of the citizen participants attended regional meetings, where they met one another and the event organizers. Most also took part in an Online Parliament that generated the principal proposals that the Canberra meetings discussed.

[8] Details available online through event sponsor, the new Democracy Foundation (http://www.newdemocracy.com.au).

[9] The Old Parliament room used for small group discussions had once served as the dining room for the Australian parliament, and its acoustics were designed to afford its guests a modicum of privacy in their conversations. In effect, this meant that it was sometimes impossible to sort out a participant's utterance from the surrounding background noise.

[10] The basis for the analysis can be found in Raudenbush and Bryk (2002, Chapter 12). We are thankful to Lesa Hoffman for making SAS code available for this type of analysis (http://psych.unl.edu/psycrs/945/12b_Changes_in_Nesting_Example.pdf).

[11] The distributions for all three tasks are positively skewed, with skewness = 0.80, 0.92, and 1.30, for the three tasks in order. Some participants spoke quite frequently; for example, one member contributed 485 speaking turns during Task 2.

[12] When using group averages, instead of grand means, similar results are obtained.